How Many ChatGPT Conversations Are Used for Training? The Math
With default settings, one chat per day means 365 conversations in OpenAI's training dataset. We calculated how much of your data is actually being used.
👁 116 reads
ChatGPT’s conversation history works on two tracks, and that’s the source of most confusion. The first layer is the simple chat log in the sidebar - what you see when you log in. The second layer is Memory, facts about you extracted from conversations that the model adds to context in future chats. And there’s a third thing that gets talked about least: the default opt-in consent for using your content to improve models.
In the English interface, this toggle is called “Improve the model for everyone” and sits in Settings, under Data Controls. It’s turned on by default for free, Plus, and Pro accounts. It’s not enabled for Teams, Enterprise, or API traffic - OpenAI describes this in their help center.
This asymmetry is crucial and few people catch it. A company that buys Team licenses for thirty people has training disabled automatically. The same company where thirty people use private Plus accounts and paste contract fragments has training enabled for all thirty. The price difference between Plus and Team today is in practice around ten to fifteen dollars monthly per person. That surcharge buys primarily a change in the default privacy setting.
The calculation: one chat daily means 365 conversations in training data
We did the simplest possible calculation, because nobody provides exactly these numbers. Assume a user who uses ChatGPT moderately - one session per day, six exchanges in it, each exchange averaging 40 words on their side. This is a very conservative assumption, because people paste entire paragraphs into ChatGPT, not sentences.
Six exchanges times 40 words equals 240 words daily. Times 365 days gives 87,600 words per year. For comparison: that’s roughly the length of a medium-sized novel or three standard master’s theses. At an average reading speed of 200 words per minute, reading through your own year of chat history would take over seven hours non-stop. And this is the material that, with default settings, is available for training models.
The second calculation concerns tokens, because that’s the unit engineers think in. For Polish text, we estimate roughly 1 token per 0.55-0.6 words, meaning Polish text is more expensive for the tokenizer than English. 87,600 words therefore equals approximately 150,000 tokens annually from one moderate user. Someone who uses ChatGPT for work - pasting code, emails, document fragments - produces that much in two or three weeks.
How many users never touched these settings? Around 90 percent
Here we need to calculate based on two independent data points, because OpenAI doesn’t publish statistics on toggle clicks. First data point: scale. OpenAI reported in 2025 over 700 million weekly users, and Poland consistently ranks in the second ten countries by traffic to chatgpt.com. A conservative estimate for Poland is several million people monthly - let’s take 5 million as a working number, because that’s roughly what you get scaling traffic according to Poland’s share of European internet usage.
The second data point is user behavior with default settings. Research on the so-called default effect in digital products has for years given convergent results: the percentage of people who independently change a default privacy option falls in the 5-15 percent range. The lowest values concern settings hidden two levels deep in menus, and the “Improve the model for everyone” toggle is exactly that - you need to enter Settings, then Data Controls.
Taking the middle of this range, 10 percent, it works out that approximately 4.5 million Polish ChatGPT users never touched the default training consent. Multiplying this by the 87,600 words per person per year calculated above, we get an order of magnitude that’s impressive: nearly 400 billion words of Polish-language user text per year. This is material incomparable to anything that can be collected from the Polish internet, because these are conversations about specific problems, written in natural language, often with professional context.
We note where this calculation is fragile: 5 million is an estimate, not a number confirmed by OpenAI, and not every conversation actually enters the training set - the company applies filtering and deduplication. But the order of magnitude doesn’t change, because an error of half in any of the assumptions still gives hundreds of billions of words.
What GDPR actually provides, and what it doesn’t
The legal basis OpenAI uses for training is legitimate interest under Article 6(1)(f) GDPR. This is important because under this basis you have the right to object under Article 21 - and in practice, clicking the toggle in Data Controls is exactly exercising that right. So this isn’t the producer’s grace, but implementation of an entitlement from the regulation.
Italy’s regulator Garante was the first in Europe to strike at OpenAI, imposing a 15 million euro fine in December 2024. The justification concerned among other things the lack of proper legal basis for processing data for training and the lack of age verification mechanism.
“OpenAI processed users’ personal data to train ChatGPT without first identifying an appropriate legal basis and violated the principle of transparency and related information obligations toward users.”
- statement from Italian Garante per la protezione dei dati personali, December 2024
The procedural problem is that OpenAI has its European headquarters in Ireland, so from February 2024 the role of lead supervisory authority passes to Ireland’s DPC, and other national offices become concerned authorities. In practice this means a complaint filed in one EU country goes to Dublin, and the decision comes back after months or years.
Two clicks worth doing today, and a third nobody talks about
Specifically and in order, because this is the part you can do while reading this paragraph. Click your initials in the upper right corner, choose Settings, then Data Controls. Switch “Improve the model for everyone” to off - from this moment new conversations don’t go to training, though history remains visible to you. In the same section is Memory: you can view the list of remembered facts about you in “Manage” and delete individual entries or all of them. This is worth doing once, because memory collects things you forget you wrote.
The third thing is Temporary chat mode, the icon next to model selection. Such a chat doesn’t save to history, doesn’t feed memory, and doesn’t enter training - it’s the equivalent of an incognito window. We use it whenever we paste anything with others’ data: a contract fragment, a candidate’s CV, a message from a client. And one more thing that gets lost in discussions about toggles: disabling training doesn’t mean content isn’t stored. OpenAI keeps conversations for 30 days for abuse monitoring purposes, and from 2025 there’s also the matter of court orders in US copyright disputes that periodically forced retention of data longer.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →For whom this isn’t a matter of hygiene but legal obligation
If you work with client, patient, or job candidate data, pasting it into ChatGPT on a free account with default settings is transferring personal data to an entity with which you don’t have a data processing agreement under Article 28 GDPR. This isn’t theoretical risk - it’s a ready charge during an audit. A law firm, clinic, accounting office, or HR agency needs a Team or Enterprise account, where training is disabled by default and OpenAI signs a DPA. The surcharge of around ten to fifteen dollars monthly per person is the cheapest element of compliance you can buy today.
A separate matter concerns parents and schools. The training toggle is hidden in the same place on a teenager’s account, and its default value is identical. If a child uses ChatGPT for homework and types in the teacher’s name, school name, and description of their own problems - and they do, because that’s what natural conversation looks like - all of this with default settings is material for improving the model. Two minutes with settings on a child’s account are worth more here than an hour of conversation about internet dangers.
Facts checked September 26, 2026. Sources: OpenAI - How your data is used to improve model performance, OpenAI Privacy Policy, Garante statement on 15 million euro fine, OpenAI - usage scale data, GDPR, Articles 6 and 21. We tested this in Polish. The estimate of Polish user numbers and the percentage not changing settings are our calculations based on public traffic data and literature on the default effect - we provide them as an order of magnitude, not as numbers confirmed by OpenAI.
More on this from QUANTSLANT
benchmarks·Wan 3.0 vs Wan 2.2: Same Prompts, 8 Clips Side by SideWe tested Wan 3.0 against Wan 2.2 with four identical prompts: eight clips, physics, audio, generation time, and cost - $0.43 versus two cents.